The Alignment Problem: Machine Learning and Human Values

Christian’s The Alignment Problem opens with a set of historical anecdotes that inform the contemporary concerns in AI research. These anecdotes raise several questions, posed by the author directly and indirectly, such as “What are the ethical implications of AI utilization?,” “What are the ways in which human and machine learning intersect?,” and “What are the unintended consequences of using predictive systems and of technological advancement in general?”

The story of Walter Pitts that Christian opens the book with serves as a metaphor for the contingent pathways that innovation can follow. While Pitts, a prodigious yet formally uneducated young man, significantly contributes to the early development of AI systems, he never does so from the position of an insider, and he’s also not widely recognized for his contribution. Christian uses the incident to illustrate a broader motif within AI development: the unexpected paths that technological advancements can take. Similarly, the creation of neural networks, initially inspired by the workings of the human brain, has led to developments that their creators could not have foreseen, such as the emergence of biased outcomes in systems like word2vec. These narratives point to the fact that technology often evolves in directions that its creators neither predict nor control, highlighting significant Ethical Implications of AI Use in contemporary contexts, especially concerning the responsibility and foresight that creators and consumers have in the development and usage of AI systems.

The analogy Christian draws between the logical model that Walter Pitts and Warren McCulloch developed and the neural network (described in the Prologue) highlights the interconnections between human cognitive processes and machine learning algorithms. Pitts’s insights into the logical structure of neural networks prefigure modern AI’s attempts to mimic aspects of human thought:

It was already known by the early 1940s that the brain is built of neurons wired together, and that each neuron has “inputs” (dendrites) as well as an “output” (axon). When the impulses coming into a neuron exceed a certain threshold, then that neuron, in turn, emits a pulse. Immediately this begins to feel, to McCulloch and Pitts, like logic: the pulse or its absence signifying on or off, yes or no, true or false (2).

The subsequent development of this AI system, described in the Introduction, is the ability of Google’s word2vec to draw analogies between cities and countries. These examples point to The Intersection of Human and Machine Learning by illustrating machines’ growing capability to perform tasks that require a form of semantic understanding, traditionally thought to be a uniquely human trait. The examples also put into question the boundaries between human and machine cognition, exploring how machines might not only replicate but also extend human intellectual capacities. However, as the narrative suggests, this blending of human and machine intelligence brings with it complex challenges, particularly when machines reflect and amplify existing human biases, such as in the use of COMPAS in the United States’s criminal justice system.

A key idea in both Christian’s Prologue and Introduction is the ethical implications of AI technology. As the narrative transitions from historical anecdotes to contemporary issues, such as the discovery made by the Microsoft Research team of the gendered bias of word2vec and the racial bias of the COMPAS system, Christian’s focus turns to the ethical challenges that arise when AI systems learn from data sets that reflect societal prejudices. The examples of AI use in life-altering decisions and key societal contexts point to the urgent requirement for transparency, accountability, and fairness in AI applications. Christian’s narrative bridges the gap between individual technological developments and their broader social impacts, urging a proactive approach to the ethical management of AI. Outlining the issue as “the alignment problem” (13)— defined as the effort of aligning human values with the development and implementation of AI—Christian sets out to explore various AI technical and intellectual advancements while also providing a critical examination of the societal responsibilities that accompany these technological powers. As he states, the stakes of AI usage are high, evoking enthusiastic Interdisciplinary Approaches to AI Development and Implementation:

In reaction to this alarm—both that the bleeding edge of research is getting ever closer to developing so-called ‘general’ intelligence and that real-world machine-learning systems are touching more and more ethically fraught parts of personal and civic life—has been a sudden, energetic response. A diverse group is mustering across traditional disciplinary lines. Nonprofits, think tanks, and institutes are taking root. Leaders within both industry and academia are speaking up, some of them for the first time, to sound notes of caution—and redirecting their research funding accordingly. The first generation of graduate students is matriculating who are focused explicitly on the ethics and safety of machine learning. The alignment problem’s first responders have arrived at the scene (15).

As AI continues to evolve and integrate into various aspects of human life, Christian introduces the readers to the context and connections between human and machine learning, as well as the ethical considerations essential to navigating the future of AI alignment with human values.